Methods for joint modelling of multiple longitudinal and survival outcomes

When we want to predict the risk of a person developing a disease, or experiencing some other kind of medical event, we build a risk prediction model. The risk prediction model is used to inform someone about their future risk of disease and/or to make treatment decisions about the most appropriate way to manage their disease risk.
Usually a risk prediction model uses recent measurements of factors that are known to influence disease risk. We are exploring whether previous measurements of these risk factors, which are commonly stored in electronic medical records, can give additional information and help us to estimate the disease risk more precisely.
In this research programme we are exploring statistical methods which allow past measurements to be used in risk prediction, and developing statistical methods to determine whether a personalised screening schedule could help identify those at high disease risk.
One application of these methods is in cardiovascular risk prediction. This work will have important implications in the management of the risk of cardiovascular disease by improving the accuracy of cardiovascular risk prediction and developing a more individualised approach to the scheduling of cardiovascular risk assessments.
We will explore the potential of using data from electronic health records, such as records from GP practices, in this research. We anticipate this type of data will bring many challenges because the data were not originally collected for research purposes, and are considered “big data” because of the number of individuals involved.

In many clinical applications we are interested in statistically modelling more than one outcome. These might be multiple longitudinal outcomes, such as the risk factors blood pressure, total cholesterol and high-density lipoprotein cholesterol for cardiovascular disease (CVD), or they may include a time-to-event, such as the first CVD event. The aim of this research programme is to develop and explore statistical methodology for the analysis of multiple outcomes. One area of focus will be dynamic risk prediction, where an individual’s prediction for the risk of an event occurring is updated in response to new measurements made of time-varying risk predictors. Methods for modelling repeated measurements of multiple risk predictors and a time-to-event outcome include joint modelling, where all outcomes are modelled simultaneously, and landmarking, where past repeated measurements are modelled separately from future times-to-events in a two-stage approach.

A clinical application which motivates this work is cardiovascular risk prediction. We are using data from electronic health records to develop a risk prediction tool for cardiovascular disease and to provide recommendations for the optimal scheduling of cardiovascular risk assessments. By projecting the risk of cardiovascular disease into the future we will aim to determine whether a person will likely be at high risk of cardiovascular disease in the future, and thereby to obtain a personalised recommendation for the time to schedule the next risk assessment. Using data from electronic health records for clinical research can be challenging because this data has not been collected for research purposes. We will explore the impact of informative observation of risk factor measurements because these measurements may have been taken in response to the state of health of the individual. There will also be computational challenges in the application of complex statistical methods to such big data. Additional questions we will explore in this context include the improvement in predictive accuracy that could be achieved by including other aspects of the risk factor trajectory in a prediction model, such as the slope of the longitudinal trajectory, or the variability in the repeated measurements, and the use of decision theory in obtaining the personalised recommendations. This work will have important implications in the management of the risk of cardiovascular disease.

Another application which motivates this programme of research is to understand the relationship between lung function and survival in cystic fibrosis. We are using joint modelling to assess whether the effect of sex on survival in cystic fibrosis patients is mediated by lung function, and will further explore the role of body mass index.